How smarter AI helps cars see through radar interference
Jeroen Overdevest defended his PhD thesis at the Department of Electrical Engineering on June 24th.
Radar sensors play a crucial role in today's vehicles. They continuously measure the distance, speed and direction of objects around the car, helping systems such as adaptive cruise control, automatic emergency braking and collision avoidance. Unlike cameras, radar sensors continue to function well in poor weather and low-light conditions. This makes them an essential component of safe and reliable driver assistance technology.
A growing challenge: radar interference
As the number of radar-equipped vehicles increases, so does the risk of interference. When multiple radar systems operate in the same frequency range, their signals can overlap and disturb one another. This can make it more difficult for a vehicle to accurately interpret its surroundings, particularly in busy traffic situations where many radar sensors are active at the same time.
Why traditional and AI-based solutions fall short
Conventional radar processing methods are based on mathematical models and physical principles. These techniques are well understood and reliable, but they can struggle with increasingly complex interference scenarios. At the same time, researchers have explored neural networks and other AI techniques to solve the problem. While these methods can learn complex patterns from data, they often function as black boxes. They require large amounts of training data, are difficult to interpret and do not always perform well when moving from simulations to real-world situations.
The Best of Both Worlds
This research of focuses on a promising alternative: model-based deep learning.Instead of relying solely on mathematics or solely on artificial intelligence, model-based deep learning combines both approaches. Physical knowledge about radar signals is built directly into the learning algorithms, allowing the system to remain interpretable while still benefiting from the flexibility of machine learning. The result is a smarter and more efficient way to deal with interference in automotive radar systems.
Effective suppression of interference
Several model-based deep learning techniques were developed and tested to suppress both radar-to-radar interference and self-interference. The results show that these methods outperform existing approaches while often requiring less computing power and memory than conventional deep learning models. This is particularly important for automotive applications, where processing resources are limited and reliability is critical.
From simulation to reality
A common challenge for AI systems is that models trained on simulated data do not always work well in the real world. The techniques developed in this research were trained using simulated radar data but were also evaluated on real-world measurements. Their strong performance demonstrates that incorporating physical knowledge helps bridge the gap between simulation and practice. This makes the technology more suitable for deployment in future vehicles.
Towards safer and more reliable radar systems
The key conclusion of this PhD research is clear: effective and reliable suppression of interference in automotive radar applications is possible by combining classical signal processing with data-driven learning. By balancing physical understanding with the adaptability of artificial intelligence, model-based deep learning offers a practical path toward more robust radar perception. As vehicles become increasingly automated and connected, such advances will help ensure that radar systems continue to provide accurate and dependable information, even in the most demanding traffic environments.
Title of PhD thesis: . Supervisors: Dr. Ruud van Sloun, , and Dr. Alessio Filippi (NXP).